<- all tokdocs
2026.09.17@ryanlpeterman

Casey Muratori Says Conway's Law Is the Only Real Law in Software, and That It Now Applies to AI Agents

Coding
Watch on TikTok

View on TikTok ->

Muratori's version of Conway's law is a bandwidth argument, and the popular "products look like the org chart" phrasing is only a side effect of it. The clip is 75 seconds cut from a Peterman Pod episode with Casey Muratori published on September 14, 2026. Muratori is the programmer behind Handmade Hero, Granny 3D at RAD Game Tools, and the collision and world editor work on Jonathan Blow's The Witness. He also gave a full lecture on this exact topic in March 2022 called "The Only Unbreakable Law," which is where the clip's title comes from.

What Conway actually wrote in 1968

Melvin Conway published "How Do Committees Invent?" in Datamation in April 1968, after Harvard Business Review rejected it. The sentence people quote is this one: "organizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations." On his own site, Conway states it as "any organization that designs a system will produce a design whose structure is a copy of the organization's communication structure."

Notice what the original sentence is about. It names communication structure, not reporting hierarchy. Muratori is making the same correction in the clip when he says "products look like the org chart" is a downstream consequence rather than the rule itself. Two teams under different VPs who talk constantly are one communication unit. One team on paper split across two time zones that sync weekly is two.

The bandwidth framing is the useful part

Muratori's phrasing in the clip is more mechanical than the usual version. His condition is that communication between workers is slower than the internal computation ability of each worker. Once that holds, splitting the problem buys parallelism, and the price is that design iteration across the split gets slower than iteration inside either half.

That framing makes the law falsifiable in a way the org chart version is not. It predicts the boundary hardens in proportion to the communication cost. It also predicts no penalty when the workers can talk as fast as they think, which is why a solo programmer or a single team in one room does not pay it. The clip does not give the derivation. The full lecture does, and so does the podcast episode.

The agents extension is Muratori's conjecture, not a result

The clip slips "or in the modern parlance, let's say agents" into the definition without flagging that this is new territory. The full episode transcript on developing.dev is more careful. Muratori says there: "I suspect it will also become true for AIs and things like that, where it's like, you will want to partition these problems in ways that respect that line drawing."

"I suspect" is the operative phrase. I found no study testing Conway's law on multi-agent LLM systems. The mechanism is at least plausible, since agents passing summaries through a context window have far less bandwidth between them than a single agent reasoning inside one context. Treat it as a prediction worth testing rather than a finding.

The empirical record behind the law

Muratori calls it a law without citing evidence in the clip. The evidence exists and it is reasonably strong.

Nagappan, Murphy and Basili studied Windows Vista at Microsoft Research and the University of Maryland, covering 3,404 binaries and over 50 million lines of code. Their organizational structure metrics predicted failure-prone binaries with 86.2% precision and 84.0% recall, against 78.6% and 79.9% for code churn. Organizational metrics beat code complexity, code coverage, code dependencies and pre-release defect counts too.

MacCormack, Baldwin and Rusnak tested the mirroring hypothesis by pairing commercial products against open source products that do the same job. In every pair they examined, the product from the loosely coupled organization was significantly more modular. Their paper ran in Research Policy in October 2012.

What you can do about it

The practical response has a name the clip never mentions. Thoughtworks calls it the inverse Conway maneuver: structure your teams to match the architecture you want, then let Conway's law build that architecture for you. It has been on their Technology Radar for years and it is the operating premise behind the Team Topologies book.

Wikipedia's summary of the disagreement is worth keeping in mind. People argue about whether the effect is good or bad, and about which direction causality runs. If architecture also shapes how teams end up communicating, then the inverse maneuver is pushing on a feedback loop rather than a one-way lever. That is a reason to expect it to be slow and messy in practice.

Key Takeaways

  • Conway's 1968 sentence is about communication structure. The org chart version is a rough proxy that fails whenever real communication patterns differ from the reporting lines.
  • Muratori's condition is specific: the law bites when communication between workers costs more than computation inside a worker. No bandwidth gap, no penalty.
  • The cost you lock in is iteration speed across the boundary, which is why bad partitions show up as designs that never quite harmonize rather than as outright bugs.
  • The Vista study found organizational metrics predicted failure-prone binaries at 86.2% precision and 84.0% recall, ahead of churn, complexity, coverage and dependency metrics.
  • The extension to AI agents is unverified. Muratori says "I suspect" in the full episode, and I found no published study testing Conway's law on multi-agent systems.
  • The venue for Muratori's 2022 lecture is also unverified. Secondary sources say it was given to a school of informatics in Madrid, and I could not confirm that against a primary source.
  • The inverse Conway maneuver is the standard counter-move, and it assumes causality runs from org to architecture, which is itself contested.

Resources

Published September 17, 2026. Writeup generated from a favorited TikTok.